EEG Based Functional Brain Network Analysis and Classification of Dyslexic Children During Sustained Attention Task

Reading is a complex cognitive skill that involves visual, attention, and linguistic skills. Because attention is one of the most important cognitive skills for reading and learning, the current study intends to examine the functional brain network connectivity implicated during sustained attention in dyslexic children. 15 dyslexic children (mean age 9.83±1.85 years) and 15 non-dyslexic children (mean age 9.91±1.97 years) were selected for this study. The children were asked to perform a visual continuous performance task (VCPT) while their electroencephalogram (EEG) signals were recorded. In dyslexic children, significant variations in task measurements revealed considerable omission and commission errors. During task performance, the dyslexic group with the absence of a small-world network had a lower clustering coefficient, a longer characteristic pathlength, and lower global and local efficiency than the non-dyslexic group (mainly in theta and alpha bands). When classifying data from the dyslexic and non-dyslexic groups, the current study achieved the maximum classification accuracy of 96.7% using a k-nearest neighbor (KNN) classifier. To summarize, our findings revealed indications of poor functional segregation and disturbed information transfer in dyslexic brain networks during a sustained attention task.


I. INTRODUCTION
D EVELOPMENTAL Dyslexia (DD) is defined as a neuro-developmental reading disorder with difficulty in learning to read or write even with average intelligence [1].With an incidence rate of 5-12% around the globe, DD is one of the most common learning disabilities among developing children.Several studies have stated impaired phonological processing is the cause of reading disabilities in dyslexia, but also there are studies reporting attention deficit and reduced functional connections in the reading network areas of the brain as one of the causes of reading disabilities in DD [1], [2].Previous studies [3], [4] have shown that reading disorder and attention disorder co-occur as attention is one of the prime cognitive abilities necessary for acquiring reading skills at developing stages.Therefore, any deficit in attention would affect the development of children's reading skills.Shaywitz and Shaywitz [5] explained the role of attention in reading and claimed that the attention deficit possibly causes reading difficulties.Also, many previous studies reported attentional deficits in dyslexic children compared to typically developing children [2], [6].With the evidence of attentional impairment in dyslexic children [2], [7], a study on investigating the attention-task related changes in the brain network connectivity of dyslexic children is missing.
In recent times, electroencephalogram (EEG) recordings are widely used to study the functional brain networks during task-related activity and rest [8], [9].EEG is the recording technique of the brain's electrical activity in a non-invasive way and provides high temporal resolution.The present study employed graph theory to study the brain network characteristics of dyslexic and non-dyslexic children using EEG recordings.Graph theory offers a potential mathematical tool to mimic and explore the complex system of brain networks [10], [11].Graph theory models the human brain as a complex network structure with nodes and edges.Nodes signify the brain areas or regions or electrode locations, and the edges represent the statistical relationship between brain areas.It has been primarily applied to illustrate the local and global communication in brain networks.Many studies have found altered network metrics in the brain's pathological condition using graph analysis [11], [12].Therefore, we have used graph theory to study the brain network characteristics in dyslexic children while performing sustained attention tasks.Recently, Dushanova et al. [13] have measured the changes in functional connectivity (FC) in dyslexic children before and after visual training using graph theory to study the effect of remedial training in the dyslexic children brain networks.The post visual training results showed more segregated network in dyslexics similar to the controls.Taran et al. [6] performed seed to voxel based FC analysis in dyslexic children using functional magnetic resonance imaging (fMRI) scans and observed reduced FC in attention related areas in brain during reading task.Although the studies in recent times have studied the functional connections in dyslexic brain, the major work were focused on resting state connectivity analysis [9] and reading task connectivity analysis [6].But with the increased evidence for attentional impairment in dyslexic children, a study on analyzing the brain networks during an attention task could provide better understanding of the brain dynamics related to attentional skills.Therefore, the present study motivated to examine the changes in the functional network connectivity measures in dyslexic children while performing sustained attention task.The sustained attention is the capacity of an individual to focus attention on particular task for a longer period.Sustained attention is one of the prime skills for reading where the individual needs to maintain attention during the process.The present study selected sustained attention (over other attention types such as selective, divided, alternating, and focused attention) as it has been suggested as a possible predictor of development of language and reading skills in children [14].
Advances in computational methods and machine learning algorithms greatly impacted neuroscience, particularly in classifying and identifying neurological disorders at early stages.Currently, the diagnosis of dyslexia primarily depends on the questionnaire-based behavioral analysis, which requires an experienced psychologist and assistance is needed from the parents to help psychologists in identification [15].The main drawback of the traditional method is that the dyslexic condition is identifiable only after specific years of acquiring education [16].Moreover, the conventional methods do not classify the subject based on the actual brain differences but instead based on behavioral assessments.So the present study is interested in using a machine learning (ML) algorithm to classify the dyslexic and non-dyslexic subjects using graph features derived from EEG data.Tamboer et al. [16] used structural magnetic resonance imaging (MRI) scans of 22 dyslexic and 27 without dyslexic students to train an support vector machine (SVM) classifier for classification and obtained a classification accuracy of 80%.Also, in the same study, Tamboer et al. [16] used another sample set of 840 young adults without dyslexia to test the trained SVM classifier and obtained a performance accuracy of 59%.Plonski et al. [17] performed multi-parameter ML methods using structural data of the dyslexic population and received above-chance accuracy in classifying the person with dyslexia against the control group.These studies demonstrate the potential use of ML algorithms combined with neuroimaging data to identify dyslexia.Based on this, the present study applied the ML methods to functional brain network measures during attention task to identify dyslexic conditions and to study ML techniques' efficacy in classifying dyslexia from non-dyslexic groups.The present study applied two classifiers-SVM [18] and k-nearest neighbor (KNN) classifier [19], which are most commonly used for classification in biological data.In summary, the present study goal was to study the changes in the brain network connectivity measures while performing sustained-attention task and to classify the dyslexic children using ML algorithm.This paper is the first to study the functional brain network connectivity during sustained attention task in dyslexic children to the authors' best knowledge.

A. Subject Selection and Task Details
A group of 15 dyslexic children (11 males and 4 females) and 15 non-dyslexic children (11 males and 4 females) have participated in the present study.Dyslexic group children were selected from special schools, whereas the non-dyslexic group children are from primary schools with good academic records.Children from both groups have been assessed with the Specific Learning Disability-Screening Questionnaire [20], Phonological awareness test (PAT) [21] and Child Behavior Checklist [22].Raven's progressive matrices test was used to evaluate the non-verbal IQ of the children in both groups.A school-based psychologist performed all the assessments with the presence of the children's parents.All participated children were right-handed and had no symptoms of other neurological and behavioral disorders.The children were made to sit relaxed on a chair during the recording procedure and asked to perform a visual continuous performance task (VCPT) -sustained attention task on a 15-inch laptop.The task trials consist of animal, plant, and human pictures and the trials were presented in four different pairs, i.e., animal-animal, animal-plant, plant-plant, and plant-human, as shown in Fig. 1.The task stimuli were presented for 100 ms and each trial consisted of two stimuli presented at an interval of 1100 ms and the time interval between trials was 3100 ms.Total of 60 trails were presented and children were instructed to press the key (right arrow) only for trail with animal-animal stimuli and ignore other stimuli.The task measures such as reaction time, omission error, and commission error were calculated for each presented trial.The descriptive statistics showing mean group differences and standard deviation (SD) (tested with Mann-Whitney U test, see section II-G Statistical analysis for more details) in age, IQ and test scores of the children from both groups were given in Table I.

B. EEG Recording
The children's brain activity was recorded using a 19-channel EEG device (RMS Maximus 32) with a 256 Hz Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

TABLE I AGE, IQ AND TEST SCORES OF SUBJECTS
sampling rate.EEG electrodes were placed on children's head scalp according to 10-20 EEG standards.Signals were recorded from prefrontal (Fp1 and Fp2), frontal (F4, F3, F7, and F8), central (C4 and C3), parietal (P4, P3), temporal (T5, T6, T4, and T3), occipital (O2 and O1), and midline (Pz, Fz, and Cz) brain regions with respect to a common reference electrode placed at cheek bone area.The contact impedance of 5k was maintained between skin and electrode during electrode placement for better signal conduction.The EEG device was enabled with a notch filter (50 Hz) to eliminate the 50 Hz interference during the recording.The recording protocol was 3 minutes baseline (eyes open) recording and 3 minutes task performance followed by 3 minutes rest period (see Fig. 1).The present study analyzed only the baseline and task data from the recording and the differences between baseline and task (baseline-task) is reported.This research study was approved by the Institutional Ethical Committee, NIT Raipur (NITRR/IEC/2019/02), and all parents have signed in the consent form along with the children.The whole EEG recording took place in the school with the presence of psychologist, children's parents, and medical staff (nurse) to ensure the children's comfort.The procedures were elaborated to children and their parents before the recording.

C. EEG Preprocessing and Feature Extraction
EEG signals are needed to be properly denoised at its preprocessing stage since the EEG signals are expected to be contaminated with noises (electrical noise and eyeblink) [23], [24].First, the recorded EEG signals are re-referenced to common average and the re-referenced EEG data was used for further analysis.At the preprocessing stage, a moving average smoothing filter (width size of 5, triangular window) was used to remove the high-frequency components present in the recorded signals (cut-off frequency of 50 Hz).The smoothing filter filtered out the high-frequency noises, but it did not affect the removal of ocular artifacts (eyeblink and eyeball movement).So, a wavelet denoising technique was further used to reduce the eyeblink and eye movements from the recorded EEG signals [24].Thresholding-based wavelet denoising was employed using Daubechies (db9, level 6) as mother wavelet and Stein's unbiased risk estimate (SURE) thresholding algorithm was used to filter out ocular artifacts and to get noise-free EEG as discussed in [24].The Wavelet packet decomposition (WPD) method was employed to extract EEG bands from the preprocessed EEG signal [25].The different EEG sub-bands (delta, theta, alpha and beta) were extracted at different decomposition levels are shown in Table II.

D. EEG Coherence
The present study used spectral coherence method to estimate the FC [10], [26].The spectral coherence of EEG data at 19 electrode locations was calculated in order to get 19 × 19 FC matrixes.The present study used coherence for FC analysis due to its clinical importance and wide use in clinical diagnosis and therapeutic interventions [27], [28].Coherence (C x y ) of signal x and, in general, denotes the degree of association between two signals in the frequency domain, as described in (1), with values ranging from 0 to 1.As a result, a 19 × 19 undirected weighted coherence matrix was created, which was then employed as the FC matrix for network analysis.
P x y -cross power spectral density of signal x and y P x x and P yy -power spectral densities of signal x and y respectively.

E. Network Analysis
The present study applied graph theory to model the brain as a network with nodes and edges.The node represents the brain regions/electrode locations, and the edges are the connections between nodes.In the present study, 19-channel EEG electrode locations were classified as nodes, and the coherence between 19-channel EEG data was specified as edges.Network measures such as characteristic pathlength, global efficiency, clustering coefficient, local efficiency and small-worldness were measured for both the baseline and task conditions.
1) Clustering Coefficient (C l ): In a network, the C l describes a node's ability to form a cluster and calculates the number of connected triangles in a network.The C l of a network considered to be a valuable measure in understanding the brain network association and calculated as mentioned in (2).
Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
The brain networks with high C l are considered to be largely efficient networks at local level [10], [26].
where, k i = j∈N w i j is the weighted degree between node i-j, t w i = 1 2 j,h∈N (w i j w i h w j h ) 1 / 3 is the number of triangles around node i, w i j is the weight between node i-j, n represents number of nodes and N represents set of nodes.
2) Characteristic Pathlength (L p ): The path length is the distance between any two nodes in a network and the shortest pathlength denotes the minimum distance from a node to other node in a network.The L p is the average shortest distance between all pair of nodes in a network and calculated as mentioned in (3).L p measure defines how well the information gets transferred within the networks [10], [26].
where d i j = a uv ∈gi w ←→ j f (w uv ) is the shortest pathlength between i-j nodes.
3) Global Efficiency (E g ): In general, the network efficiency refers to the efficiency of information exchange in a network.The E g measures the efficiency of information transfer among entire nodes in the network and calculated as mentioned in ( 4).E g measure quantifies the complete information processing and parallel transfer of information among scattered subnetworks [10], [26].
The E l provides a measure of how effectively information gets transferred among the first neighbors of a given node in the network [10], [26].In network analysis, higher E l values indicated the high local segregated neural processing.The E l of a network was calculated as in ( 5) 5) Small-World Network (S w ): The small-world network (S w ) defines a brain network organization with highly clustered nodes (high C l ) and less distance/length between nodes (low L p ) [10], [26], and is calculated as normalised C l divided by normalised L p as mentioned in (6).
The normalized C l and normalized L p of a network are obtained by normalizing the C l and L p of original matrix using C r and L r (clustering coefficient and pathlength of random matrix) of 100 random networks.The generated random network matrices are similar to original networks with randomly shuffled edge weights from the original matrix [29].All the network metrics calculations were performed using the brain connectivity toolbox (BCT) in MATLAB for graph analysis [26].

F. Task Induced Changes in Network Measures
The present study calculated the task induced changes in the network measures in order to evaluate the changes in the network measures due to task performance.The task-induced changes in the network measures ( C l , L p , E g, E l and S w ) were calculated as the difference in the values of network measure during the task and baseline periods, as described in ( 7), ( 8), ( 9), (10), and (11) )

G. Statistical Analysis
The non-parametric method was used in statistical analysis since the measured network parameters did not match the normality test assumptions using the Shapiro-Wilk and Kolmogorov-Smirnov tests.The alpha value for this experiment was set to 0.05.The same group analysis across condition was evaluated using post-hoc Wilcoxon sign test.For multiple comparisons (4 bands X 5 features), Bonferroni corrections were used, and the new value of alpha was adjusted to 0.0025 (0.05/20).The Mann-Whitney U test was used to look for significant differences between groups.IBM SPSS version 20 was used to conduct all statistical analyses.

H. Classifier and Performance Evaluation
The network measures during task ([C l ] task , [L p ] task , [E g ] task , [E l ] task and [S w ] task ) and task induced changes in network measures ( C l , L p , E g, E l and S w ) were used for classification.This study uses a linear SVM classifier and a KNN classifier (k=3) for classification.The performance of classifiers was evaluated using a 5-fold and leave one subject out cross-validation (LOSOCV) technique.The subject data was divided into training and testing subsets for all iterations during cross validation (CV).In 5-fold CV the data was divided into 5 folds and each fold holds 20% of the total data.So in 5-fold CV, 4 folds (80%) were used for training and the remaining one fold (20%) was used for testing the classifier performance and it is repeated with different test fold for all CV iterations.The classification accuracy at all 5-fold CV iterations were calculated and the mean accuracy of 5-fold CV was reported.In LOSOCV, one subject data was used for testing and the remaining 29 subjects data was used for training the classifier and the model performance were calculated.The performance parameters such as accuracy (Acc), sensitivity (S e ), specificity (S p ) and area under the curve (AUC) were calculated from the confusion matrix.The confusion matrix holds the information of the actual class and predicted class.The dyslexics were classified as a positive class, whereas the non-dyslexics were classified as a negative

TABLE III TASK PERFORMANCE MEASURES
class.As a result, the true positive signifies correct dyslexic group classification, while the true negative denotes correct non-dyslexic group identification.MATLAB R2019a was used to carry out the classification methods.The entire method used in this investigation is depicted in Fig. 2.

A. Task Performance
The task measures suggest that the dyslexic group has a higher omission error (p<0.0001) when comparing the dyslexic and non-dyslexic groups (see Table III) indicates the dyslexic group is easily distracted and not fully focused on the items presented.Furthermore, the dyslexic group had considerably higher commission errors (p<0.0001),indicating that they were impulsive when executing the task, and that this caused them to respond to a non-target stimulus.Although there were no significant differences in reaction time (RT) between groups, the dyslexic group had a higher mean RT than non-dyslexics.The dyslexic children performed poorly on the cognitive task that needed constant attention, as evidenced by the task measures.

B. Network Measures in Dyslexic Group During Task and Baseline Condition
Table IV compares within-group variations in the dyslexic group across conditions and the highly significant p-value was highlighted in bold text.In the dyslexic group, there was no significant variation in C l value across conditions.In the dyslexic group, the delta band L p value during the task was considerably lower (p<0.001)than the baseline condition.At the theta band, the E g and E l were found to be considerably lower (p<0.001)during the task condition compared to the baseline condition in the dyslexic group.We found no significant differences in S w values in the dyslexic group across conditions.

C. Network Measures in Non-Dyslexic Group During Task and Baseline Condition
Table V compares within-group differences across conditions in the non-dyslexic group.In the non-dyslexic group, the C l measure was significantly higher during task conditions at theta (p<0.001),alpha (p<0.001), and beta (p=0.002)bands compared to baseline.During the task condition, the L p measure was found to be considerably lower in the delta (p=0.002) and theta (p<0.001)bands than in the baseline condition.When compared to the baseline condition, the E g was significantly higher (p<0.001)under the task condition at delta, theta, and alpha band.Only during the task condition was the E l value found to be considerably higher (p=0.001)than in the baseline condition.When compared to the baseline period, the S w value for non-dyslexic was considerably higher (p=0.001)during task time at theta and alpha bands.

D. Network Measures During Task Condition Between Groups
The differences between the groups under the task condition are shown in Table VI.When the C l value during Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

TABLE V NETWORK MEASURES OF NON-DYSLEXIC GROUP ACROSS CONDITION TABLE VI COMPARISON OF NETWORK MEASURES DURING TASK BETWEEN GROUPS
task condition was compared to the non-dyslexic group, the dyslexic group exhibited considerably lower C l values in the theta (p=0.002) and alpha (p<0.001)bands.During the task, the L p value in the dyslexic group was considerably higher than in the non-dyslexic group at the theta band (p<0.001).When comparing the task conditions between groups, we found no significant differences in the values of E g

TABLE VII COMPARISON OF NETWORK MEASURES DURING BASELINE BETWEEN GROUPS
and E l .In the theta and alpha bands, the S w value in dyslexics was substantially lower (p<0.001)than in non-dyslexics.

E. Network Measures During Baseline Condition Between Groups
At the beta band, the C l value in the dyslexic group was found to be substantially higher (p<0.001)than in the non-dyslexic group during the baseline condition (see Table Table VII).At the theta and alpha bands, the L p value in dyslexics was found to be lower (p<0.001)than in non-dyslexics.During the baseline condition, the E g value in dyslexics was considerably higher (p<0.001)than non-dyslexics in the theta, alpha, and beta bands.During baseline, dyslexics had considerably higher E l values in the theta (p<0.001) and alpha (p=0.002)bands than nondyslexics.When comparing dyslexics to non-dyslexics, the S w value in theta band was found to be much lower (p<0.001)for dyslexics.Furthermore, we did not notice a small-work network in the dyslexic group during task and baseline time because the mean S w values in both conditions were less than 1.

F. Task Induced Changes in Network Measures
Fig. 3 shows the distribution of C l values of dyslexic and non-dyslexic groups using split violin plot.In the violin plot, the solid line represents the group mean, while the dotted line represents the squared error.When comparing the C l values, we found substantial group differences in the beta band.The mean C l was found to be considerably lower in the dyslexic group compared to the non-dyslexic group, indicating that the dyslexic group had lower C l values during task time.
Using a split violin plot, Fig. 4 depicts the distribution of L p values in the dyslexic and non-dyslexic groups.When comparing the L p values between groups, significant differences in theta and alpha bands were found.When compared to the non-dyslexic group, the dyslexic group's mean L p was considerably higher in the theta and alpha bands.Fig. 5 shows the distribution of E g values of dyslexic and non-dyslexic groups using split violin plot.When comparing the E g values between groups, significant differences were seen in the theta and alpha bands.When comparing the dyslexic group to the non-dyslexic group, the mean E g was found to be considerably lower in the theta and alpha bands.Fig. 6 shows the distribution of E l values of dyslexic and non-dyslexic group using split violin plot.Only the theta band showed significant changes between groups.The mean theta E l in the dyslexic group was considerably lower than in the non-dyslexic group, with E l values decreasing from baseline during the task in the dyslexic group.
The comparison of small-world metrics between groups is shown in Table VIII.When comparing the two groups, we found substantial variations in S w at theta and alpha bands.At theta and alpha bands, the S w was found to be considerably lower in dyslexics compared to non-dyslexics.

G. Classifier Performance
Through statistical analysis (Mann-Whitney U test), a total of 13 features were found as statistically distinct features.With the help of Waikato Environment for Knowledge Analysis (WEKA) software, feature selection was done using the correlation attribute ranking approach [25].The top 5 ranking features (Beta_ C l, Alpha_ L p , Theta_ E g , Theta_ S w and Theta_ L p ) were chosen for classification based on the correlation values of attributes with class (correlation coefficient > 0.7).So, for classification analysis using SVM and KNN classifiers, a 30 × 5 (Subjects X Features) feature Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

TABLE IX CLASSIFICATION PERFORMANCE EVALUATION MEASURES
space was created.Table IX shows the output of two classifiers as well as the performance of the classifiers for two typical validation techniques.It was discovered that the KNN classifier (with k=3) had the highest accuracy of 96.7% in classifying the dataset with 5-fold cross validation and also had the highest area under curve (AUC) when compared to other classifiers.We also achieved 93.3% accuracy using KNN (k=3) using the LOSOCV approach, with the greatest AUC of 93%.The linear SVM classifier showed accuracy of 93.3% with 5-fold and 90% with LOSOCV.

IV. DISCUSSION
Reading is a complex cognitive process, and it also requires other cognitive skills mainly attentional skills.Previous researches have shown that any impairment in attention mechanism may cause possible difficulties in attaining fluent reading skills and the identification of attention role in dyslexic children could help the intervention of dyslexia with new approaches [5], [30].As a result, the present study is interested in investigating the differences in the attention skills in dyslexic and non-dyslexic children.The present study used sustained attention task (VCPT) to study the differences in the brain networks.The task performance results showed significant differences in commission and omission error measures.The dyslexic group showed significantly higher errors while performing sustained attention tasks than nondyslexic children.Significant differences were not observed in RT measures between groups, but the mean value of RT was high in dyslexic group.The high commission error in the dyslexic group children showed their poor inhibitory control and the high omission error reflects the attention failure compared to non-dyslexic group.The task performance results revealed that the dyslexic group children had difficulty in performing sustained attention task which reflects reduced attention skills in dyslexic group.The current study's task performance findings are consistent with the behavioral findings of [31] who observed reduced sustained attention and inhibitory control in dyslexic children assessed using Conners' continuous performance task.Though behavioral results suggest poor sustained attention in dyslexic group, it needs more converging details that could be demonstrated using neuroimaging studies.As a result, the current research looked at the brain networks involved in sustained attention in dyslexic children.Many previous studies have linked dyslexia with reduced connectivity in the left hemisphere of the brain, in particular areas related to reading skills [32].The reading areas in the brain are also accountable for other cognitive processes such as attention, working memory, calculation, and fact retrieval [33].As a result, any dysfunction caused in the reading area of the brain might also affect the other cognitive functions associated with it and leads to comorbid conditions.Considering this, a whole-brain network analysis could help to understand the underlying neural basis of attentional skills in dyslexic children.The network analysis was performed using graph theory methods.To investigate the network topology in dyslexic and non-dyslexic children, graph network parameters such as characteristic pathlength, global efficiency, local efficiency, clustering coefficient, and small-world were determined.

A. Differences in Network Organization of Dyslexics
Generally, brain networks possess three different forms of network characteristics, namely-regular network (high C l and high L p values), random network (low C l and low L p values), and the small-world network (high C l and low L p values) [34].A normal brain activity generally possesses a small-world network efficiently which maintains high local segregation and global integration for efficient information processing in a network with cost effective manner [35].Functional segregation and integration are the two prime factors in a small-world network organization of a functional brain.The present study evaluated the functional segregation and integration using C l , L p , E l , and E g measures.The functional segregation of a network defines how well the network forms a cluster locally in a network for information processing and the functional integration defines how well the nodes in the network are connected for an efficient information transfer [36].A normal brain dynamically reconfigures based on the task demands and the optimal balancing between segregated and integrated states in the functional brain network is associated with cognitive abilities and abnormal conditions [36].Previous studies have shown altered structural and resting networks in dyslexic children.Lou et al. [37] analyzed the white matter connectivity in dyslexic children and reported less functional integrated and reduced functional segregation among distant brain regions.Lou et al. [37] further stated that poor literacy skill in dyslexic children emerges due to less integrated and segregated brain networks.Gonzalez et al. [9] found substantial group differences in the theta band and concluded that dyslexic children had a less efficient brain network during rest.The present study results provide clear significant differences in the networks of dyslexic and non-dyslexic children.We observed significant differences within the group across the conditions.In dyslexic group, we observed significantly reduced L p during task compared to baseline at delta band and the efficiency measures (E g and E l ) were significantly low during task time compared to baseline at theta bands.In contrast, the non-dyslexic group showed significant increase in network efficiencies (E g and E l ) during task time compared to baseline and significant reduction in L p during task time at delta and theta bands.Also, the C l measure found to be significantly high during task at delta, theta and alpha bands compared to baseline.The within group analysis across condition indicates clear significant differences in the network measures in both the groups while performing task and at rest.Further, we compared the network measures during task between groups and observed significant differences in C l , L p and S w measures during task time and when comparing the baseline period, we observed significant differences in all network measures.The C l and S w measures found to be significantly low during task and L p measure was significantly high during task in dyslexics compared to non-dyslexics mainly at theta and alpha bands.When comparing the baseline between groups we observed increased C l (in beta band), E g (in delta, theta and beta bands), E l (in theta and alpha bands) and decreased L p and S w .
With the clear differences between the groups when comparing the task and baseline period, the present study interested in calculating the task induced changes in the network measures in order to evaluate changes in the network measures due to the effect of task performance.The task induced changes were evaluated as the differences in the task measures and baseline measures and the group differences were studied.At the alpha, theta, and beta bands, we observed decreased C l values from baseline during task time (negative C l ) in the dyslexic group, whereas the non-dyslexic group showed raised C l values during task time.The dyslexic group's lower C l during task time reflects a less segregated network and slower information processing.Similarly, during task time at theta band, the dyslexic group had lower E l values than the nondyslexic group, whereas the non-dyslexic group had higher E l values.The decrease in E l in a dyslexic group also shows an ineffective information flow transfer in a local network.At theta and alpha bands, the dyslexic group's L p values were considerably greater during task time than the baseline (positive L p ) compared to the non-dyslexic group.Usually, the longer pathlength induced during attention/working memory task performances are unstable and expensive metabolically than the shorter pathlength which forms locally in a network [13].The high L p values in dyslexic network during task time clearly indicates the longer pathlength in a network for interaction which further suggests the unstable links and reduced integration in the dyslexic group.
When compared to the non-dyslexic group, the dyslexic group's E g values decreased during task time in the theta and alpha bands.This demonstrates the decreased efficiency of information transfer in a dyslexic group's whole network.During the task and baseline periods, we observed a small-world network for non-dyslexics in the theta, alpha, and beta bands (S w >1 for all conditions).The non-dyslexic group's S w was much higher in the theta, alpha, and beta bands.We did not observe the small-world network in dyslexic groups in any bands during task time or baseline period.The absence of small-world network in dyslexic group shows the altered network topology with imbalance in network local processing and integration.In summary, the absence of small-world network and reduction of C l , E g , E l with increased L p confirms the disrupted brain network organization in dyslexic brain and suggest a disparity in the functional integration and segregation.In contrast, the non-dyslexic group showed small-world network and had higher values of C l , E g , E l with reduced L p showed the brain network maintains its right balance of functional segregation and integration state with respect to task demands and aids in better task performance.Also, it should be highlighted that in the theta band, we observed highly segregated (high C l ) and integrated network (high L p ) with high efficiencies (high E g and high E l ) in non-dyslexic group.The dyslexic group, on the other hand, had a less segregated and inefficient network organization in theta band.Generally theta oscillations are associated with information processing and transfer among brain cortical areas and also responsible for distant communication among brain areas during cognitive processing [38].So the inefficient network of dyslexic group in theta band could be the reason for the attentional processing deficit and leads to poor task performance in dyslexic group.Our findings of theta band abnormality in dyslexic children also reported by previous researchers in dyslexic population.The present study findings are in line with the findings of [9] wherein, significant changes in the theta band during the resting time and reduced network integration and communication between nodes in dyslexic children was observed.Further, the present study findings are similar to a previous study on structural brain networks of Chinese dyslexic children [39], where the authors stated significant changes in the brain topology in Chinese DD children compared to normal children.

B. Classification
The present work also studied the efficiency of ML algorithms such as SVM and KNN in classifying the datasets of dyslexic and non-dyslexic children.The two standard validation protocols (5-fold and LOSOCV) were used to evaluate the classifier performance.The classification results showed the highest mean accuracy of 96.7% (AUC of 96%) for KNN (k=3) classifier and 93.3% (AUC of 93%) for linear SVM classifier with 5-fold validation.Using LOSOCV method, we observed satisfactory mean accuracy of 93.3% (AUC of 93%) for KNN (k=3) classifier and 90% for linear SVM classifier.Comparing the sensitivity and specificity, we observed mean sensitivity of 93% and specificity of 100% in KNN (k=3) classifier.The present study findings showed the efficiency of SVM and KNN in classifying the dyslexic group and non-dyslexic group using network based EEG features.Table X shows the comparison of present study accuracy with the available literatures.Compared to previous studies, the present study achieved the highest Acc of 96.7% with network measures derived using EEG.
The present study holds certain limitations which need to be considered while understanding the data and findings.The samples included in this study is limited due to practical restrictions however the samples included in this study is homogenous considering the criteria of inclusion-exclusion and the sample size (n=30) could be related with similar neuroimaging research on children with dyslexia [45].The present study results may prone to volume conduction effect as this is one of the drawbacks in EEG based connectivity studies [46].For future work, authors suggest a study to compare the efficiency of connectivity and non-connectivity measures in dyslexia detection.

V. CONCLUSION
The present study examined the functional brain networks in dyslexic and non-dyslexic children during sustained attention task.When comparing the dyslexic and non-dyslexic groups, the task performance findings revealed poor inhibitory control and attention failure in the dyslexic group.The variations in network measures between groups revealed indications of disrupted brain networks in the dyslexic brain while performing sustained-attention task, as evidenced by poorer functional integration and reduced global information transmission in the brain networks.Furthermore, the present study achieved the highest mean classification accuracy of 96.7% using graph features for the classification of dyslexic children.Our findings show that network measures are effective in understanding the underlying neurological basis of the attentional mechanism and that network measures may be used to classify dyslexia.

Fig. 2 .
Fig. 2. Proposed methodology for extraction of connectivity features and classification of dyslexic children.

TABLE II EEG
BANDS AND ITS DECOMPOSITION LEVELS